Cloud Workload Analytics for Real-Time Prediction of User Request Patterns
Hengjian Wang, John Pannereselvam, Lu Liu, Yao Lu, Xiaojun Zhai, Haider Ali · 2018
Energy-aware datacenter operation is gaining attention in the recent years, having witnessed the increased deployments and usage of cloud datacenters resulting in significant amount of carbon footprints to directly affect environmental sustainability. Since providers tend to record user behaviors and cloud jobs profiles, analyzing cloud trace logs can provide insightful inferences to move towards datacenter sustainability. This paper investigates the trace logs of real datacenter workload arrival trend in order to characterize the resource request trend of users. Furthermore, anticipated resource request trends in terms of the requested level of CPU and memory resources are forecasted. A novel prediction model, named LSTMtsw, is proposed based on recurrent neural networks of deep learning, which exploits historical data in combination with the current job arrival trend to predict the future resource request trend of users. By the way of constructing sliding windows of historical data, the proposed model effectively reduces the number of iterations involved in the prediction process, and predicts the future trend with reduced error percentage. In addition, the proposed LSTMtsw model integrates an effective mechanism for controlling the prediction effect to achieve energy conservation without degrading the service quality. Experiments conducted on 600 hours of cloud trace logs demonstrate the efficiencies of the proposed model in outperforming the existing BPNN and LSTM models in terms of the achieved prediction accuracy whilst predicting the future resource demands of users.